Florida Carcass Recovery Record · 1974–2023
Every Florida manatee carcass recovered over fifty years, matched to the ocean, weather, seagrass, coastline and vessel traffic around it — and modeled to show where death risk concentrates, and which driver is doing the work in each place.
Cause groups follow the FWC necropsy field. Undetermined and perinatal deaths are shown throughout but are not modeled as driver-specific outcomes.
01 · Observed mortality
Each point is one recovered carcass, placed where it was found. Filter by cause, year and season — the map, the county ranking and the timeline all respond together. Carcass recovery is presence–only data: it records where animals were found, which is why the risk model later on corrects for recovery effort.
47 of Florida's 67 counties have recorded manatee mortality. Ranked by count; selecting one filters every chart on this page and zooms the map to it.
Shaded band = the 7.5-mile coastal analysis domain. Points jitter slightly at identical coordinates.
Top 12 counties in the current filter
Stacked by cause. Note the 2021 Indian River Lagoon starvation event.
Recoveries peak in late winter. Cold stress is almost purely a December–March event (96% of its deaths); watercraft deaths are close to aseasonal.
02 · The evidence base
For every one of the 16,023 carcasses, roughly forty environmental covariates were measured inside an event-specific buffer around the recovery point, in the weeks and the year that death occurred. Six data families, six different instruments — each chosen because it carries a distinct mechanism.
Every carcass carries two footprints. A fixed 7.5-mile buffer (~458 km²) gives a constant, comparable window across all fifty years. A dynamic buffer (78–2,827 km², median 314 km²) scales per event, tightening around locations where the animal is unlikely to have drifted far. Every covariate is computed for both, so any result can be checked against the other geometry. A third zone — a statewide 7.5-mile coast-to-inland band of 39,384 km² — supplies the land-cover-change backdrop that the offshore buffers cannot.
Results shown on this page use the fixed 7.5-mile buffer.
03 · Effect sizes
Cause-specific point-process models were fitted with a recovery-effort offset, so what they describe is risk of death, not merely of being found. Each bar is the change in relative risk across the observed range of that covariate, holding the others and space and year fixed. A multiplier of 1× means no effect.
Log scale. Whiskers are 95% confidence intervals; a bar crossing 1× is not distinguishable from no effect.
Discrimination under two independent hold-outs.
Spatial-block CV holds out whole regions, so nearby autocorrelated points cannot leak between training and test. Temporal hold-out withholds the most recent years. Both matter: a model can look excellent under random CV and still fail on a new coastline or a new decade.
Risk rises 13.1× (95% CI 6.3–27.5) from the least- to the most-trafficked water in the record. Traffic is the strongest non-geographic predictor of any cause, and it predicts the right cause — it does far less work in the cold-stress model.
Cold-stress deaths are preceded by markedly colder air: median coldest weekly temperature −0.6 °C versus +1.7 °C for other causes, and 57% versus 31% experienced a sub-freezing week.
In the fitted risk model, however, the isolated air-temperature term is 0.91× (0.62–1.34) — it crosses 1. Winter cold is near-uniform across the peninsula in any given year, so once the space-by-year smooth is in the model there is little independent variation left for the temperature term to explain. The exposure contrast is the evidence here; the partial effect is not.
Higher seagrass fraction is associated with higher, not lower, mortality risk (2.09× overall). This is habitat overlap, not harm: manatees die where manatees are, and they are where the forage is. It is a reminder that these are risk-of-occurrence surfaces, not causal dose–response curves.
The single largest term in every model is distance from the coastline (0.01–0.03×, i.e. risk collapses offshore and inland). That is partly biology and partly recovery effort — carcasses are found from shorelines and boats. The effort offset absorbs some of this; it does not absorb all of it.
04 · Forecast
Predicted relative mortality risk on a 5 km coastal grid, with recovery effort held constant so the map shows intrinsic risk rather than where people happen to look. Each cause has its own geography — which is the point: a single "manatee risk map" would average away the very differences that determine which intervention works where.
05 · Priorities
Taking the top 15% of cells on the all-cause risk surface and joining any that sit within 12 km of each other yields ten coherent hotspots. Every carcass within 10 km of a hotspot is then attributed to it, which lets each one be profiled against the statewide record. The useful comparison is not which hotspot is biggest — it is which cause runs above its statewide rate there, because that is what names the intervention.
Cause share divided by its statewide share. 1.0× means the hotspot looks like Florida as a whole; red means over-represented, blue under-represented. Cells with fewer than 15 deaths of that cause are left blank.
Read across a row to see what makes a place distinctive; read down a column to see where one mechanism concentrates. The statewide baseline is watercraft 19.1%, perinatal 18.2%, red tide 16.2%, cold stress 8.0%, other human-related 2.0%, flood gate / lock 1.8%, undetermined 34.8%.
Median conditions in the buffers of carcasses attributed to each hotspot, shown against the range across all ten.
Three things are common to all ten; four things separate them into distinct intervention classes.
Median distance from the coastline is under 1 mile in nine of the ten. These are lagoons, bays, river mouths and passes — not open coast. Every hotspot is a place where manatee habitat and human access are compressed into the same thin band of water.
Undetermined share runs from 20% at Crystal River to 75% in Florida Bay. Even in the best-diagnosed hotspot, one death in five has no assigned cause — so every enrichment figure below is a statement about diagnosed deaths, and the diagnosis rate itself varies with how remote the place is.
No hotspot has one mechanism above 2× except Biscayne Bay. Everywhere else at least two causes are active, which means single-lever interventions will cap out well short of eliminating a hotspot.
Caloosahatchee & Estero Bay (1.63×) and Tampa Bay to Charlotte Harbor (1.56×) are the only two hotspots enriched for red tide, and together they hold 4,776 deaths. They are also the two most recent-weighted large hotspots. Nothing on the Atlantic side looks like them.
Biscayne Bay & Miami canals runs flood-gate and lock deaths at 8.6× the statewide rate — an order of magnitude beyond any other enrichment on this page. It also has the highest vessel traffic and the most impervious surface. It is the only hotspot whose leading cause is fixed by retrofitting hardware rather than by managing behavior or habitat.
Cold stress is enriched at the Southeast ICW (1.81×), the Indian River Lagoon (1.24×) and the Ten Thousand Islands (1.22×). Crystal River has by far the coldest air (median coldest week −2.7 °C) yet the lowest cold-stress share — its springs work. That contrast is the clearest natural experiment in the dataset for what a reliable warm-water refuge is worth.
Florida Bay (75% undetermined) and the Big Bend (49%, n = 61) have no cause above its statewide rate. They are remote, undeveloped and barely trafficked. Treating them as intervention targets would be a mistake; the actionable gap there is recovery and necropsy capacity, not management of any hazard.
Derivation. Cells of the all-cause risk surface above the 85th percentile (relative risk ≥ 0.038) were grouped by single-linkage clustering at a 12 km threshold; clusters of at least six 5 km cells (150 km²) were kept, and one cluster with only 11 attributed deaths was dropped as too thin. Carcasses within 10 km of any cell in a cluster are attributed to it; a carcass can in principle fall inside two adjacent hotspots, though in practice the ten are well separated.
The traffic caveat. Median vessel traffic at Caloosahatchee & Estero Bay reads 36,000 transits against 723,000 at Tampa Bay to Charlotte Harbor — a twentyfold gap — yet their watercraft shares are close (17.3% and 21.2%). That gap is almost certainly AIS coverage, not boating: AIS carriage is mandated for larger and commercial vessels, so the small recreational craft that dominate the Caloosahatchee and Estero Bay are largely invisible in it. Do not read the traffic column as a boating-intensity ranking between hotspots. It is reliable where the fleet is large and commercial and it understates dense small-boat estuaries.
What this section cannot tell you. Enrichment says a cause is over-represented among diagnosed deaths in a place; it does not say the cause is increasing, nor that intervening on it will avert a proportional number of deaths. The risk surfaces are relative and effort-corrected, but they are not a causal effect estimate and are no substitute for FWC's own designations.
06 · Attribution
Counting deaths after a hurricane tells you almost nothing — deaths happen anyway. The test here compares observed deaths inside each storm's wind swath, in the eight weeks after landfall, against what the baseline risk model expects for that place and season. Across 104 storms since 1974: 1,700 observed against 1,246 expected — +454 excess deaths, a rate ratio of 1.36.
Storms with the largest excess. Bar length is excess deaths; the label shows the rate ratio.
Peak wind against excess deaths, one dot per storm.
Eta (2020) was a Category 4 that contributed +117 excess deaths — roughly four times baseline. Irma (2017) and Ian (2022), both far stronger at peak, contributed +10 and +20. Track, timing and how much manatee habitat the swath covered matter more than intensity.
07 · Context
Manatees are confined to shallow, warm, fresh-water-adjacent coastal margins — precisely the strip that people build on. Across the statewide 7.5-mile coastal band, mean impervious surface rose from 7.2% in 1986 to 10.9% in 2023. Inside the buffers where carcasses were actually recovered, it rose from 9.6% to 13.9%: deaths concentrate in the more-developed part of an increasingly developed coast.
Annual NLCD fractional impervious, 1986–2023. Hover any year for its development-intensity breakdown.
Impervious surface is the annual land-use signal here because categorical land-cover classes were available for 1985 only in this data mirror. It is a good proxy for what matters to a manatee — marinas, ramps, hardened shoreline, and the runoff that feeds algal blooms — but it is a proxy, not a direct measure of any of them.
08 · Open data
Everything behind this page is downloadable. Filtered exports respect the controls you set in section 1, so you can pull exactly the slice you were looking at. Carcass records originate with the Florida Fish and Wildlife Conservation Commission's Marine Mammal Pathobiology Laboratory; covariates are derived from the public sources listed in the methods below.
09 · How to read this
This is a decision-support tool built on presence-only data. That imposes real constraints, and the honest thing is to state them alongside the maps rather than in a footnote nobody opens.
Carcass recoveries are presence-only: we know where deaths were found, but there is no denominator telling us where deaths could have occurred and didn't. To build a risk surface you need that denominator, so the design is a use–availability point process: the 16,023 recoveries are the cases, and 80,000 background quadrature points spread across the coastal domain and across time supply the availability. A down-weighted Poisson generalised additive model fitted over cases plus background approximates a log-Gaussian Cox process (Warton & Shepherd 2010; Renner et al. 2015).
Each model carries a two-dimensional spatial smooth, a smooth of year, smooths of the covariates, and an
offset(log effort) term for recovery effort. Without that offset the model forecasts
detection — where carcasses get found — rather than where manatees die. Models are fitted
separately per cause, because averaging a winter thermal process together with a summer collision process
produces a surface that describes neither.
1. Recovery location is not death location. Carcasses drift. The fixed 7.5-mile buffer is wide partly to absorb this, which in turn dilutes genuinely local processes such as a single busy channel.
2. Vessel traffic is a 2015–2025 mean applied to a 1974–2023 record. AIS does not exist for most of the study period, so traffic enters as chronic exposure — how boated a place is in general — not as a time-matched annual value. A watercraft trend over time cannot be attributed to traffic with these data.
3. Land-cover classes are 1985-only in this mirror. Annual impervious surface substitutes for annual class change; true annual class transitions need the full Annual NLCD product.
4. Concurvity among the spatial terms is high (0.94–0.98 for seagrass and coastal distance against the spatial smooth). Individual coefficients are therefore not safely interpretable in isolation; the partial effects and the isolated-driver refits shown here are the defensible reading.
5. 35% of deaths are cause-undetermined (5,575 of 16,023). If undetermined deaths are not a random sample of causes — and decomposed carcasses in remote water probably are not — cause-specific surfaces are biased toward whichever causes get diagnosed.
6. Seagrass associations run positive. Risk rises with seagrass because manatees are where forage is. Nothing on this page should be read as "seagrass is dangerous".
Zoom and pan. Every map supports scroll-wheel zoom, drag-to-pan, double-click to zoom in (shift‑double‑click to zoom out), pinch on touch, and the + / − / □ controls in the corner. Selecting a county from the search frames the map on it automatically.
The four basemaps are all drawn from data, not from tiles. Chart is coastline plus the 7.5-mile analysis domain. Counties uses US Census TIGER/Line cartographic boundaries, shaded by each county's share of the mortality record. Seagrass is the project's own 100 m seagrass raster (UNEP‑WCMC and FWC/SEACAR) aggregated to 2 km. Vessel traffic is the project's own 2015–2025 AIS transit composite aggregated to 3 km. Coastlines are Natural Earth 10 m.
Satellite and hybrid use a Sentinel‑2 surface-reflectance composite built for this page in Google Earth Engine: dry-season scenes from winter 2024–25 under 20% cloud, masked per-pixel for cloud, shadow and cirrus using the scene classification band, then median-reduced and rendered as true color. It is a 6×5 tile mosaic — a single whole-extent median exceeds Earth Engine's memory limit — stitched and embedded in this file as one image, which is why it still works with no network. Hybrid draws the coastline, county lines and the analysis domain over it. Imagery: Copernicus Sentinel data, processed by ESA.
Because the composite is embedded rather than streamed from a tile server, it has a fixed resolution: sharp at state and regional zoom, visibly soft if you zoom all the way into a single pass or channel. That is the deliberate trade for a page that is one self-contained file.
Thao, Z., Svenson, E., & Uelmen, J. Environmental stressors, habitat condition, and the spatiotemporal risk of Florida manatee (Trichechus manatus latirostris) mortality: a reproducible geospatial framework to inform conservation. Manuscript in preparation, Department of Population Health Sciences, School of Medicine and Public Health, University of Wisconsin–Madison.
Carcass data: Florida Fish and Wildlife Conservation Commission, Fish and Wildlife Research Institute, Marine Mammal Pathobiology Laboratory. Covariates derive from NOAA/NCEI sea-surface temperature, PRISM daily air temperature, MRLC NLCD annual fractional impervious surface and land cover, NOAA Marine Cadastre AIS vessel transits, NASA ocean-color chlorophyll and Kd490, UNEP-WCMC and FWC/SEACAR seagrass extent, and the NOAA IBTrACS best-track archive.
Please cite the manuscript rather than this page once it is published, and contact the authors before reusing the risk surfaces for management decisions — they are relative, not absolute, and are not a substitute for FWC's own designations.